DSP

Official implementation of "DSP: Dual Soft-Paste for Unsupervised Domain Adaptive Semantic Segmentation". Accepted by ACM Multimedia 2021.

Authors: Li Gao, Jing Zhang, Lefei Zhang, Dacheng Tao.

Prerequisite

CUDA/CUDNN

Python3

PyTorch==1.7

Packages found in requirements.txt

Creat a new conda environment

conda create -n dsp_env python=3.7
conda activate dsp_env
conda install pytorch=1.7 torchvision torchaudio cudatoolkit -c pytorch
pip install -r requirements.txt

Download the code from github and change the directory

git clone https://github.com/GaoLii/DSP/
cd DSP

Prepare dataset

Download Cityscapes, GTA5 and SYNTHIA dataset, then organize the folder as follows:

├── ../../dataset/
│   ├── Cityscapes/     
|   |   ├── gtFine/
|   |   ├── leftImg8bit/
│   ├── GTA5/
|   |   ├── images/
|   |   ├── labels/
│   ├── RAND_CITYSCAPES/ 
|   |   ├── GT/
|   |   ├── RGB/
...

Training and Evaluation example

Training and evaluation are on a single Tesla V100 GPU.

Train with unsupervised domain adaptation

python train.py

Evaluation

python evaluateUDA.py --model-path checkpoint.pth

Pretrained models

[Pretrained model for GTA5->Cityscapes](Link: https://pan.baidu.com/s/10adjjSXarJOvat-ibzfoLg Code: wv28).
[Pretrained model for SYNTHIA->Cityscapes](Link: https://pan.baidu.com/s/1FOXwfkihUZeWmEvh3RM15g Code: a9pj).

This model should be unzipped in the '../saved' folder.

License

The code is heavily borrowed from DACS.

If you use this code in your research please consider citing

@article{Gao_2021,
   title={DSP: Dual Soft-Paste for Unsupervised Domain Adaptive Semantic Segmentation},
   url={https://arxiv.org/abs/2107.09600},
   DOI={10.1145/3474085.3475186},
   journal={Proceedings of the 29th ACM International Conference on Multimedia},
   publisher={ACM},
   author={Gao, Li and Zhang, Jing and Zhang, Lefei and Tao, Dacheng},
   year={2021},
   month={Oct}
}